Re: [PATCH 1/3] arm64: Disable GiC priorities on Mediatek devices w/ firmware issues

[Mark Rutland]

On Fri, Oct 06, 2023 at 03:15:51PM -0700, Douglas Anderson wrote:
> In commit 44bd78dd2b88 ("irqchip/gic-v3: Disable pseudo NMIs on
> Mediatek devices w/ firmware issues") we added a method for detecting
> Mediatek devices with broken firmware and disabled pseudo-NMI. While
> that worked, it didn't address the problem at a deep enough level.
> 
> The fundamental issue with this broken firmware is that it's not
> saving and restoring several important GICR registers. The current
> list is believed to be:
> * GICR_NUM_IPRIORITYR
> * GICR_CTLR
> * GICR_ISPENDR0
> * GICR_ISACTIVER0
> * GICR_NSACR
> 
> Pseudo-NMI didn't work because it was the only thing (currently) in
> the kernel that relied on the broken registers, so forcing pseudo-NMI
> off was an effective fix. However, it could be observed that calling
> system_uses_irq_prio_masking() on these systems still returned
> "true". That caused confusion and led to the need for
> commit a07a59415217 ("arm64: smp: avoid NMI IPIs with broken MediaTek
> FW"). It's worried that the incorrect value returned by
> system_uses_irq_prio_masking() on these systems will continue to
> confuse future developers.
> 
> Let's fix the issue a little more completely by disabling IRQ
> priorities at a deeper level in the kernel. Once we do this we can
> revert some of the other bits of code dealing with this quirk.
> 
> Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
> ---
> 
>  arch/arm64/kernel/cpufeature.c | 21 +++++++++++++++++++++
>  1 file changed, 21 insertions(+)
> 
> diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
> index 2806a2850e78..e35efab8efa9 100644
> --- a/arch/arm64/kernel/cpufeature.c
> +++ b/arch/arm64/kernel/cpufeature.c
> @@ -2094,9 +2094,30 @@ static int __init early_enable_pseudo_nmi(char *p)
>  }
>  early_param("irqchip.gicv3_pseudo_nmi", early_enable_pseudo_nmi);
>  
> +static bool are_gic_priorities_broken(void)
> +{
> +	bool is_broken = false;
> +	struct device_node *np;
> +
> +	/*
> +	 * Detect broken Mediatek firmware that doesn't properly save and
> +	 * restore GIC priorities.
> +	 */
> +	np = of_find_compatible_node(NULL, NULL, "arm,gic-v3");
> +	if (np) {
> +		is_broken = of_property_read_bool(np, "mediatek,broken-save-restore-fw");
> +		of_node_put(np);
> +	}
> +
> +	return is_broken;
> +}

I'm definitely in favour of detecting this in the cpucap, but I think it'd be
better to parse the DT once on the boot CPU rather than on each CPU every time
it's brought up.

I think if we add something like:

#ifdef CONFIG_ARM64_PSEUDO_NMI
static void detect_system_supports_pseudo_nmi(void)
{
	struct device_node *np;

	if (!enable_pseudo_nmi)
		return;
	
	/*
	 * Detect broken Mediatek firmware that doesn't properly save and
	 * restore GIC priorities.
	 */
	np = of_find_compatible_node(NULL, NULL, "arm,gic-v3");
	if (np && of_property_read_bool(np, "mediatek,broken-save-restore-fw")) {
		pr_info("Pseudo-NMI disabled due to Mediatek Chromebook GICR save problem");
		enable_pseudo_nmi = false;
	}
	of_node_put(np);
}
#endif /* CONFIG_ARM64_PSEUDO_NMI */
static inline void detect_system_supports_pseudo_nmi(void) { }
#endif

... then we can call that from init_cpu_features() before we call
setup_boot_cpu_capabilities(), and then the existing logic in
can_use_gic_priorities() should just work as that returns the value of
enable_pseudo_nmi.

Note: of_node_put(NULL) does nothing, like kfree(NULL), so it's fine for that
to be called in the !np case.

Would you be happy to fold that in? I'm happy with a Suggested-by tag if so. :)

Mark