Re: [PATCH] mmc: host: meson-gx-mmc: fix possible deadlock condition for preempt_rt

[Jerome Brunet]

On Fri 25 Sep 2020 at 15:44, Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> wrote:

> On 2020-09-25 11:11:42 [+0200], Jerome Brunet wrote:
>> I'm not sure about this.
>> As you have explained on IRC, I understand that IRQF_ONESHOT is causing
>> trouble with RT as the hard IRQ part of the thread will not be migrated
>> to a thread. That was certainly not the intent when putting this flag.
>
> That is my understanding as well.
>
>> This seems pretty unsafe to me. Maybe we could improve the driver so it
>> copes with this case gracefully. ATM, I don't think it would.
>
> Running the primary handler in hardirq context is bad, because it
> invokes meson_mmc_request_done() at the very end. And here:
> - mmc_complete_cmd() -> complete_all()
>   There is a lockdep_assert_RT_in_threaded_ctx() which should trigger.
>
> - led_trigger_event() -> led_trigger_event()
>   This should trigger a might_sleep() warning somewhere.
>
> So removing IRQF_ONESHOT is okay but it should additionally disable the
> IRQ source in meson_mmc_irq() and re-enable back in
> meson_mmc_irq_thread(). Otherwise the IRQ remains asserted and may fire
> multiple times before the thread has a chance to run.

Looks like we need to do manually what IRQF_ONESHOT was doing for us :(
This brings a few questions:

* The consideration you described is not mentioned near the description
  of IRQF_ONESHOT. Maybe it should so other drivers with same intent
  don't end up in the same pitfall ?

* Why doesn't RT move the IRQ with this flag ? Seems completly unrelated
  to RT (maybe it is the same documentation problem) 

* Can't we have flag doing the irq disable in the same way while still
  allowing to RT to do its magic ? seems better than open coding it in
  the driver ?

>
> Sebastian