Re: [PATCH] acpi/rt: convert acpi_gbl_gpe_lock to raw spinlock

[Aaron Sierra]

----- Original Message -----
> From: "Thomas Gleixner" <tglx@xxxxxxxxxxxxx>
> Sent: Thursday, December 1, 2016 8:52:48 AM

> On Wed, 30 Nov 2016, Aaron Sierra wrote:
>> When testing GPE interrupts with CONFIG_PREEMPT_RT_FULL enabled, a
>> verbose WARN_ONCE message would print to the kernel log. It turned out
>> that the GPE interrupt handler was being called with local interrupts
>> enabled because acpi_gbl_gpe_lock was implemented as a spinlock_t. Full
>> preemption strips local interrupt disabling from spinlock_t operations,
>> but not for raw_spinlock_t operations.
>> 
>> This is the warning that was triggered:
>> 
>> ------------[ cut here ]------------
>> WARNING: CPU: 8 PID: 483 at kernel/irq/handle.c:149
>> __handle_irq_event_percpu+0x6f/0xcf
>> irq 33 handler irq_default_primary_handler+0x0/0xb enabled interrupts
>> Modules linked in: gpio_irq_demo(O)
>> CPU: 8 PID: 483 Comm: irq/9-acpi Tainted: G           O
>> 4.8.6-rt5-00012-geaa3b7c #6
>> Hardware name: Extreme Engineering Solutions, Inc. XCalibur4643/XCalibur4643,
>> BIOS 1-1.1.12.3_Alpha 04/29/2016
>>  0000000000000000 ffff880858f3bc10 ffffffff81219a93 ffff880858f3bc60
>>  0000000000000000 ffff880858f3bc50 ffffffff8104b84a 0000009500000000
>>  ffff880855b76880 0000000000000021 0000000000000002 ffff880856356800
>> Call Trace:
>>  [<ffffffff81219a93>] dump_stack+0x4d/0x63
>>  [<ffffffff8104b84a>] __warn+0xc0/0xdb
>>  [<ffffffff8104b8af>] warn_slowpath_fmt+0x4a/0x4c
>>  [<ffffffff810f7192>] ? path_openat+0xbf8/0xc62
>>  [<ffffffff8107d7c3>] ? handle_irq_event+0x75/0x75
>>  [<ffffffff8107d68d>] __handle_irq_event_percpu+0x6f/0xcf
>>  [<ffffffff8107d727>] handle_irq_event_percpu+0x3a/0x61
>>  [<ffffffff8107d7a3>] handle_irq_event+0x55/0x75
>>  [<ffffffff8107ff76>] handle_simple_irq+0x5c/0x92
>>  [<ffffffff81243a53>] gpe_irq_handler+0x2a/0x31
> 
> gpe_irq_handler is not in tree, so I really cannot tell what this is
> about. It looks like it does interrupt demultiplexing, which triggers the
> warning.

Thomas,
Your guess is correct, this involves a currently out-of-tree extension to
the gpio-ich driver which allows GPIO pins to be mapped to IRQs by demuxing
their corresponding GPE events.

The gpio_irq_demo module requests a GPIO pin, converts it to an IRQ using
gpio_to_irq() and registers a trivial handler against the IRQ.

> We tried to make that lock raw years ago and this results in a different
> splat: https://marc.info/?l=linux-kernel&m=132468512523207&w=2

Thanks for the background.

> So no, we are not making that lock raw again blindly just because of random
> out of tree code.

I thought that might be that case.

-Aaron S.