On 10/2/2023 3:58 PM, Stephen Boyd wrote:
Quoting Abhinav Kumar (2023-09-28 17:46:11)
On 9/27/2023 3:01 PM, Stephen Boyd wrote:
Quoting Kuogee Hsieh (2023-09-25 09:07:18)
However for external DP case, link training can not be guarantee always
success without link rate or lane being reduced as Abhinav mentioned.
In addition, CTS (compliance test) it required to complete link
training within 10ms after hpd asserted.
Is it possible to change that timeout? I have to look around for the CTS
parameters because I'm pretty confused how it can work. What do we do if
DP wakes the system from suspend and asserts HPD? We need resume time to
be < 10ms? That's not realistic.
No, the CTS doesnt say we need to finish link training within 10ms after
HPD is asserted. It says it must be completed in 10ms after
TRAINING_PATTERN_SET dpcd write.
"Wait until the Source DUT writes 00h to the TRAINING_PATTERN_SET byte
of Reference Sink DPCD Link Configuration Field to indicate the end of
the link training. Stop the link training timer. Verify that link
training completed in 10ms or less"
That needs to be done independent of HPD so we can ignore the CTS point.
Great!
I am not sure do link training at atomic_enable() can meet this timing
requirement.
Why? It's putting some time bound on link training in general to only
take 10ms, right?
Like I said, CTS is mentioning 10ms to finish link training after the
DUT writes 00h to the TRAINING_PATTERN_SET byte. So for this discussion
lets leave out CTS for now.
At least in the DP spec itself it doesn't require the link to be trained
within 10ms of HPD being asserted. Instead it simply recommends that the
OS start configuring the display promptly after HPD is asserted, e.g.
within 100ms. There's some strict timing on IRQ_HPD, so the driver must
read DPCD registers within 100ms of IRQ_HPD rising edge; maybe that is
what CTS is checking for?
TL;DR: I don't see why CTS should stop us from link training in
atomic_enable(). It would be beneficial to do so to make eDP and DP the
same. It would also help to report a drm connector being connected
_before_ link training so that userspace knows the link itself is the
bad part of the equation (and not that the DP connector looks
disconnected to userspace when in fact it really is connected and the
monitor is asserting HPD, just the link training failed).
Its the corrective action of the userspace when it finds link is bad is
the concern as I highlighted in the other response. Just reading and
resetting link_status is not enough to recover.
What needs to be done to recover? Userspace will try to set a mode on
the connector again if the link status is bad and there were some modes
available. If there are zero modes and the link is bad, then it ignores
the connector. I'm not sure what else could be done to recover besides
try again and stop trying if no modes exist.
Let me re-explain if I didnt make this clear last time.
You are right. Thats all the "userspace" can do which is basically retry
the mode. And like I said, its again only going to fail. All the
corrective actions you mentioned below like ignoring the connector
entirely or consider that the display has link training problems are not
something we decided to go with on a commercial device where we expect
things to be more reliable.
Let me re-explain what I explained in the prev response.
If driver issues hot-plug after link-training:
It would have implemented all the link training mechanisms such as
trying lower rates/number of lanes and made sure that when the usermode
queries the list of modes, only the modes which fit into the link rate
which was link trained successfully will be exposed and the chances of a
user ending up with a blank screen on connection are pretty high.
This reduces the dependency on usermodes to be smart enough to implement
such policies and we would rather not depend on those unless we have
some reference to a compositor which is more sturdy. I do not think the
CrOS code you have pointed to is more sturdy than the driver mechanism
explained above.
As opposed to this, if we just issue hotplug without any of this,
usermode does not know which mode to retry as we do not remove or edit
the mode list once link training fails.
Acting like the connector isn't connected makes the situation worse for
ChromeOS because userspace thinks there's nothing there so it can't try
to retrain the link again. Instead, userspace has to rely on the kernel
driver to train the link again. The kernel should just tell userspace
the link is bad so userspace can implement the policy to either ignore
the connector entirely or to consider it a display that is having link
training problems.
What gain will it give if it retries the same mode blindly as opposed to
the safer option I have explained above. None of the policies you have
highlighted seem like something an end user will be satisfied with.
So again, I see no reason why the kernel driver thinks it can implement
a policy to train the link before indicating the drm connector is
connected. It should stop doing that. Instead it should tell userspace
that the connector is connected and then train the link when there's a
modeset. If the modeset fails then userspace can take action to either
figure out that the link is bad, or notify the user that the cable is
bad, or to try replugging or power cycle the monitor, etc. None of that
can be done if the kernel lies about the state of the connector because
the link training failed.
Usermode is unable to take the corrective action without proper support
from the kernel like removing unsupported modes etc and I dont see other
drivers taking an action like that. Kernel is not lying. Its delaying
the status to a point where usermode can safely handle.
Please explain to me how any of the policies you have explained usermode
can take are safer and have more chance of success than what we have now.
