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

[khsieh@codeaurora.org]

On 2021-05-20 22:00, Stephen Boyd wrote:
> Quoting khsieh@codeaurora.org (2021-05-20 15:39:09)
>> On 2021-05-20 14:28, Stephen Boyd wrote:
>> > Quoting khsieh@codeaurora.org (2021-05-20 13:05:48)
>> >> On 2021-05-20 12:28, Stephen Boyd wrote:
>> >> >> 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.
>> >
>> > Agreed.
>> >
>> >>
>> >> I think you mistakenly think DP_LINK_STATUS_UPDATED is caused by
>> >> unplug
>> >> interrupt.
>> >
>> > Ok, got it.
>> >
>> >> 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.
>> >
>> > Agreed.
>> >
>> >>
>> >> 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.
>> >>
>> >
>> > I'm mostly trying to point out that the irq handling and masking needs
>> > to be done in the hard irq context and not in the kthread. It may or
>> > may
>> > not be related to this message that's printed.
>> >
>> > What happens if the hardirq is blocked by some other irq that takes a
>> > long time to process? Imagine this scenario:
>> >
>> > CPU0                                CPU1
>> > ----                                ----
>> >  really_long_other_hardirq() {
>> >                                     hpd_irq
>> >                                   hpd_irq
>> >                                   hpd low
>> >  }
>> >
>> >  dp_display_irq_handler() {
>> >
>> 
>> 
>> >    <fork things to kthread>
>> >  }
>> >
>> > Shouldn't we ignore any hpd_irq events in this scenario? And shouldn't
>> > we be disabling the hpd_irq by masking it with DP_DP_IRQ_HPD_INT_MASK
>> > when hpd goes low (i.e. DP_DP_HPD_UNPLUG_INT_MASK)?
>> 
>> 
>> 
>> 1) irq_hpd interrupt always happen before unplug interrupt
>> 2)if hdp_isr_status = (DP_DP_IRQ_HPD_INT_MASK |
>> DP_DP_HPD_UNPLUG_INT_MASK) at the time when read at
>> dp_display_irq_handler(),
>> then DP_DP_IRQ_HPD_INT_MASK will be add into evetn q first followed by
>> DP_DP_HPD_UNPLUG_INT_MASK be add into event q.
>> So that DP_DP_IRQ_HPD_INT_MASK will be executed by the event thread
>> before DP_DP_HPD_UNPLUG_INT_MASK.
>> On the other word, IRQ_HPD has higher priority over UNPLUG in the 
>> timing
>> matter.
>> By doing that we can shut down display gracefully.
> 
> Ok. So you're saying that we want to put both events on the queue
> regardless, and put IRQ_HPD there first because we want to check the
> status bit? Doesn't reading the status bit require the dongle to be
> connected though? So if an unplug came in along with an irq_hpd we may
> queue both the irq_hpd and the unplug, but when it comes time to 
> process
> the irq_hpd in the kthread the link will be gone and so trying the dpcd
> read for the link status will fail?
> 
yes,
we had a previous bug with this scenarios already.
https://partnerissuetracker.corp.google.com/issues/170598152
At this case, dongle produce two interrupts, irq_hpd followed by unplug 
immediately (not presented at isr status register at same time), at the 
time dongle unplugged form DTU.
But due to dp ctrl reset at handling irq_hpd which cause unplug mask bit 
be cleared so that unplug interrupt got lost.


I think V6 patch is good to go.


>> 
>> If you insist, at hdp_isr_status = (DP_DP_IRQ_HPD_INT_MASK |
>> DP_DP_HPD_UNPLUG_INT_MASK) case,
>> we can have only add DP_DP_HPD_UNPLUG_INT_MASK to event q only by
>> dropping DP_DP_IRQ_HPD_INT_MASK.
>> Is this will work for you?
>> 
> 
> I'm not insisting on anything. I'm just trying to think of various
> corner cases that we're not handling in the code so we don't have to
> worry about them in the future.