Re: [PATCH v4 1/2] drm/msm/dp: handle irq_hpd with sink_count = 0 correctly

From: khsieh@codeaurora.org
To: Stephen Boyd <swboyd@chromium.org>
Cc: agross@kernel.org, bjorn.andersson@linaro.org,
	robdclark@gmail.com, sean@poorly.run, vkoul@kernel.org,
	abhinavk@codeaurora.org, aravindh@codeaurora.org,
	linux-arm-msm@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH v4 1/2] drm/msm/dp: handle irq_hpd with sink_count = 0 correctly
Date: Thu, 20 May 2021 13:05:48 -0700	[thread overview]
Message-ID: <f476d82d0798e0d7eb9e12949aa2c8f1@codeaurora.org> (raw)
In-Reply-To: <CAE-0n52rBrjy-=dpqK+dae2GNk1rAaQnKqCjzdqiAoS13gHpSQ@mail.gmail.com>

On 2021-05-20 12:28, Stephen Boyd wrote:
> Quoting khsieh@codeaurora.org (2021-05-20 09:08:03)
>> On 2021-05-19 14:06, Stephen Boyd wrote:
>> > Quoting khsieh@codeaurora.org (2021-05-19 09:01:02)
>> >> On 2021-05-18 14:42, Stephen Boyd wrote:
>> >> > Quoting Kuogee Hsieh (2021-05-14 10:35:13)
>> >> >> irq_hpd interrupt should be handled after dongle plugged in and
>> >> >> before dongle unplugged. Hence irq_hpd interrupt is enabled at
>> >> >> the end of the plugin handle and disabled at the beginning of
>> >> >> unplugged handle. Current irq_hpd with sink_count = 0 is wrongly
>> >> >> handled same as the dongle unplugged which tears down the mainlink
>> >> >> and disables the phy. This patch fixes this problem by only tearing
>> >> >> down the mainlink but keeping phy enabled at irq_hpd with
>> >> >> sink_count = 0 handle so that next irq_hpd with sink_count =1 can be
>> >> >> handled by setup mainlink only.
>> >> >>
>> >> >> Changes in v2:
>> >> >> -- add ctrl->phy_Power_count
>> >> >>
>> >> >> Changes in v3:
>> >> >> -- del ctrl->phy_Power_count
>> >> >> -- add phy_power_off to dp_ctrl_off_link_stream()
>> >> >>
>> >> >> Changes in v4:
>> >> >> -- return immediately if clock disable failed at
>> >> >> dp_ctrl_off_link_stream()
>> >> >>
>> >> >> Signed-off-by: Kuogee Hsieh <khsieh@codeaurora.org>
>> >> >
>> >> > I think we want some Fixes tag. Not sure what it would be though.
>> >> >
>> >> > I also noticed that if I plug and unplug the HDMI cable from my apple
>> >> > dongle that I see this error message
>> >> >
>> >> >   [drm:dp_display_usbpd_attention_cb] *ERROR* Disconnected, no
>> >> > DP_LINK_STATUS_UPDATED
>> >>
>> >> > *ERROR* Disconnected, no DP_LINK_STATUS_UPDATED <== this is caused by
>> >> > dongle generate the second
>> >> irq_hpd with sink_count = 0 after first first irq_hpd with sink_count
>> >> =
>> >> 0. The fix is you have
>> >> set dongle to D3 (power off) state after first irq_pd with sink_count
>> >> =display_disable
>> >> 0 handled.
>> >> I have a patch fix this problem. I will merge and re submit for
>> >> review.
>> >
>> > That's good. I still don't understand how the kthread can't race with
>> > irq_hpd and hpd going low though. Userspace will have to disable
>> > thectrl_off_link_stream()).
>> > display and that could happen far later than the time that the hpd low
>> > interrupt fires and is processed. Can't hpd go high during that time
>> > and
>> > then blip before userspace notices and disables the display?
>> >
>> > Put another way, putting the dongle into D3 state may make the race
>> > window smaller, but it's not fixing the root cause of the problem which
>> > is that the kthread is running later and userspace is involved in the
>> > state of the dongle while irqs are firing. The three different contexts
>> > have to coordinate work, so it feels like a better approach would be to
>> > shut off the irq_hpd interrupt once hpd goes low for an unplug in
>> > hardirq context so that we don't have to consider the cable state or
>> > userspace changing the state after we notify it.
>> 
>> There is no race condition here.
>> The interrupts are converted into event and stored at event q.
>> event thread service event sequentially and make sure transaction had
>> been completed before service next event.
>> The first irq_hpd with sink_count = 0 is handled and this transaction
>> will not completed until user space frame work turn off display
>> (msm_dp_display_disable()).
>> After that, the second irq_hpd with sink_count will be service which
>> found that display is off so it spill out DP_LINK_STATUS_UPDATED 
>> warning
>> message and do nothing.
>> 
>> Put dongle to D3 state so that it will not issue the unnecessary 
>> second
>> irq_hpd with sink_count = 0. this will prevent the annoy but unharmful
>> DP_LINK_STATUS_UPDATED warning message.
>> Again, we can not disable hpd interrupt since dongle still attached 
>> and
>> hdmi cable can be plugged in at any instant.
>> 
> 
> Right I'm not suggesting to disable hpd interrupt, just the hpd_irq
> interrupt once an unplug irq comes in, and do that in hardirq context.
> Also, I'm suggesting that we consider unplug as a higher priority if 
> the
> hard irq handler is delayed for some reason and both an unplug irq and
> an hpd irq are pending in the hardware when the hard irq handler is
> running. Putting the dongle into D3 state won't fix these problems.

The unplug interrupt is not happen in this case since dongle still 
attached.
The unplug interrupt only happen when dongle unplugged.

I think you mistakenly think DP_LINK_STATUS_UPDATED is caused by unplug 
interrupt.
DP_LINK_STATUS_UPDATED happen is due to dongle issue two consecutive 
irq_hpd with sink_count = 0 when hdmi cable unplugged from dongle.
The first irq_hpd with sink_count = 0 is handled as expected to turn off 
display.
After that the second irq_hpd with sink_count = 0 is handled.
Since display had turned off, then there is nothing to do but spill 
DP_LINK_STATUS_UPDATED warning message.
There is no unplug (hpd become low) happen in this case since dongle 
still attached.

All interrupt (plug/irq_hpd and unplug) are required to be handled in 
the order of happening.
We can not ignore any one.
For example, you plug/unplug two different resolution monitor 
alternative to/from dongle and unplug dongle once for while.

I think the race condition you describe here all had been taken care 
with
1) convert irq into event and store at event q in order.
2) irq handled base on transaction. Next irq can be handled when 
previous irq transaction is done.