Re: [PATCH] x86/i8259: Work around buggy legacy PIC

From: Maximilian Luz <luzmaximilian@gmail.com>
To: Thomas Gleixner <tglx@linutronix.de>,
	"H. Peter Anvin" <hpa@zytor.com>, Sachi King <nakato@nakato.io>,
	Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
	David Laight <David.Laight@aculab.com>
Cc: "x86@kernel.org" <x86@kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"stable@vger.kernel.org" <stable@vger.kernel.org>
Subject: Re: [PATCH] x86/i8259: Work around buggy legacy PIC
Date: Mon, 17 May 2021 21:25:51 +0200	[thread overview]
Message-ID: <b509418f-9fff-ab27-b460-ecbe6fdea09a@gmail.com> (raw)
In-Reply-To: <87a6otfblh.ffs@nanos.tec.linutronix.de>

On 5/17/21 8:40 PM, Thomas Gleixner wrote:
> Max,
> 
> On Sat, May 15 2021 at 00:47, Maximilian Luz wrote:
>> I believe the theory was that, while the PIC is advertised in ACPI, it
>> might be expected to not be used and only present for some legacy reason
>> (therefore untested and buggy). Which I believe led to the question
>> whether we shouldn't prefer IOAPIC on systems like that in general. So I
>> guess it comes down to how you define "systems like that". By Tomas'
>> comment, I guess it should be possible to implement this as "systems
>> that should prefer IOAPIC over legacy PIC" quirk.
>>
>> I guess all modern machines should have an IOAPIC, so it might also be
>> preferable to expand that definition, maybe over time and with enough
>> testing.
> 
> I just double checked and we actually can boot just fine without the
> PIC even when it is advertised, but disfunctional.
> 
> Can you please add "apic=verbose" to the kernel command line and provide
> full dmesg output for a kernel w/o your patch and one with your patch
> applied?

I don't actually own an affected device, but I'm sure Sachi can provide
you with that.

As far as we can tell, due to the NULL PIC being chosen nr_legacy_irqs()
returns 0. That in turn causes mp_check_pin_attr() to return false
because is_level and active_low don't seem to match the expected values.
That check is essentially ignored if nr_legacy_irqs() returns a high
enough value. I guess that might also be a firmware bug here? Not sure
where the expected values come from.

Due to this, mp_map_pin_to_irq() fails with -EBUSY which causes
acpi_register_gsi() to fail. That fails in acpi_dev_get_irqresource(),
which causes the IRQ resource to be marked as disabled.

Down the line, this then causes platform_get_irq() to return -EINVAL,
because the IRQ we're trying to get has the IORESOURCE_DISABLED bit set.

Sachi can probably walk you through this a bit better as she's the one
who tracked this down. See also [1, 2] and following comments.

Regards,
Max

[1]: https://github.com/linux-surface/linux-surface/issues/425#issuecomment-835309201
[2]: https://github.com/linux-surface/linux-surface/issues/425#issuecomment-835261784