Re: linux-next: Tree for Aug 31 (new arm, arm64, s390 failures)

[Guenter Roeck <linux@roeck-us.net>]

On 08/31/2015 12:13 PM, Marc Zyngier wrote:
> On Mon, 31 Aug 2015 11:57:14 -0700
> Guenter Roeck <linux@roeck-us.net> wrote:
>
>> Hi Marc,
>>
>> On 08/31/2015 11:26 AM, Marc Zyngier wrote:
>> [ ... ]
>>>
>>> Actually, the kernel dies because of this:
>>>
>>> commit adaac459759db4a1fd35baddbe47bac700095496
>>> Author: Markus Pargmann <mpa@pengutronix.de>
>>> Date:   Sun Aug 30 09:33:53 2015 +0200
>>>
>>>       regmap: Introduce max_raw_read/write for regmap_bulk_read/write
>>>
>>>       There are some buses which have a limit on the maximum number of
>>>       bytes that can be send/received. An example for this is
>>>       I2C_FUNC_SMBUS_I2C_BLOCK which does not support any reads/writes of
>>>       more than 32 bytes. The regmap_bulk operations should still be able
>>>       to utilize the full 32 bytes in this case.
>>>
>>>       Signed-off-by: Markus Pargmann <mpa@pengutronix.de>
>>>       Signed-off-by: Mark Brown <broonie@kernel.org>
>>>
>>> which never considers bus to be NULL in __regmap_init. With the
>>> following patch applied, I can boot to a prompt:
>>>
>>>   From 031eae5a1b34f952ba3dcaecb4eb4ec9d3bda352 Mon Sep 17 00:00:00 2001
>>> From: Marc Zyngier <marc.zyngier@arm.com>
>>> Date: Mon, 31 Aug 2015 19:16:16 +0100
>>> Subject: [PATCH] regmap: Fix max_raw_read/write handling when bus is NULL
>>>
>>> Commit adaac459759d ("regmap: Introduce max_raw_read/write
>>> for regmap_bulk_read/write") added new fields to regmap_bus
>>> and started using them in __regmap_init, but failed to
>>> consider the case where bus would be NULL, like in the
>>> vexpress-syscgf case. The box (actually its qemu version)
>>> ends up dying painfully.
>>>
>>> Fix it by testing bus before doing anything else.
>>>
>>> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
>>
>> Yes, that fixes the vexpress failures.
>>
>> Tested-by: Guenter Roeck <linux@roeck-us.net>
>>
>> However, I still need to revert your patches to get my realview-pb-a8
>> and realview-eb tests to work again.
>>
>> I am a bit concerned about the use of of_address_to_resource(),
>> which can return an error, leaving cpu_res undefined. Also, if
>> both CONFIG_OF and CONFIG_ACPI are not configured, supports_deactivate
>> is set to true by default. This is the configuration used in my
>> failing tests.
>>
>> With that in mind, I ran a simple test.
>>
>> -static struct static_key supports_deactivate = STATIC_KEY_INIT_TRUE;
>> +static struct static_key supports_deactivate = STATIC_KEY_INIT_FALSE;
>>
>> Bingo, problem solved. Can you try to find a clean solution ?
>
> Yeah, I just came to the same conclusion, and to the following patch:
>
>  From 059bc80a4fc433dda99d75a712f30a77ce4964bc Mon Sep 17 00:00:00 2001
> From: Marc Zyngier <marc.zyngier@arm.com>
> Date: Mon, 31 Aug 2015 20:00:35 +0100
> Subject: [PATCH] irqchip/gic: Fix EOImode settiing non-DT/ACPI systems
>
> Non-DT/ACPI systems call directly into the GIC driver at init time.
> Since 0b996fd35957 ("irqchip/GIC: Convert to EOImode == 1"), this
> breaks old non firmware-driven platforms, as the driver only
> works out the capability of the platform on the DT/ACPI paths.
>
> Fix this thinko by forcing EOImode==0 on non-DT platforms,
> which are not capable of supporting a hypervisor anyway.
>
> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>

Yep, that fixes the problem.

Tested-by: Guenter Roeck <linux@roeck-us.net>

Thanks,
Guenter