3.16rc3 multiplatform, Armada 370 and IOMMU: unbootable kernel

[Thomas Petazzoni]

Dear Ezequiel Garcia,

On Mon, 7 Jul 2014 20:37:58 -0300, Ezequiel Garcia wrote:

> It seems bus_register_notifier() is been called for platform and pci devices
> with the *same* notifier block. Haven't looked close enough, but you mentioned
> that could cause trouble?
> 
> This patch fixes the issue here:
> 
> diff --git a/arch/arm/mach-mvebu/coherency.c b/arch/arm/mach-mvebu/coherency.c
> index 477202f..2bdc323 100644
> --- a/arch/arm/mach-mvebu/coherency.c
> +++ b/arch/arm/mach-mvebu/coherency.c
> @@ -292,6 +292,10 @@ static struct notifier_block mvebu_hwcc_nb = {
>  	.notifier_call = mvebu_hwcc_notifier,
>  };
>  
> +static struct notifier_block mvebu_hwcc_pci_nb = {
> +	.notifier_call = mvebu_hwcc_notifier,
> +};
> +
>  static void __init armada_370_coherency_init(struct device_node *np)
>  {
>  	struct resource res;
> @@ -427,7 +431,7 @@ static int __init coherency_pci_init(void)
>  {
>  	if (coherency_available())
>  		bus_register_notifier(&pci_bus_type,
> -				       &mvebu_hwcc_nb);
> +				       &mvebu_hwcc_pci_nb);
>  	return 0;
>  }
>  
> Paolo, can you apply it and confirm it fixes the problem?
> 
> Greg, can you confirm using the same notifier block pointer
> for two different bus types makes the bus notifier go nuts?

Looking at how notifier_chain_register() is implemented (which gets
called by bus_register_notifier() ->
blocking_notifier_chain_register()), I indeed don't see how a single
'struct notifier_block' can be registered on multiple notifier chains,
so I believe your patch is correct.

Thanks for the investigation!

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com